Using squeeze-and-excitation blocks to improve an accuracy of automatically grading knee osteoarthritis severity using convolutional neural networks

In this paper, we investigate the effect of squeeze-and-excitation blocks on improving the classification quality of osteoarthritis using convolutional neural networks of the ResNet and DenseNet families. We show that the use of these blocks improves the quality of osteoarthritis classification according to the Kellgren-Lawrence scale by 1 - 3 % without a significant modification of the model structure. We also demonstrate that combining the 0 and 1 classes of the Kellgren-Lawrence scale into one class allows one to increase the accuracy of osteoarthritis grading by 12.74 %, without losing significant information about the disease. The best final accuracy attained was 84.66 % when using an ensemble of three convolutional networks with the DenseNet-121 architecture using squeeze-and-excitation blocks, which significantly exceeds the performance of the existing state-of-the-art. The obtained results can be used both for a preliminary automatic diagnosis and as an auxiliary tool.